AI and Copyright: Did the Model Steal the Books, or Pay for Them?

"Use every means necessary to take to court the artificial intelligence companies that plunder our work."
(from the resolution filed by members of the SACD, spring 2026)

In the spring of 2026, three stories are moving forward at the same time, and almost nobody looks at them together.

Members of the SACD (the Société des auteurs et compositeurs dramatiques, the French collecting society that manages the rights of tens of thousands of theatre, film and television authors) are circulating a resolution asking the 25 June general assembly to take generative AI providers to court. To be discussed at all, it needs 4,000 votes, and under article 34 III of the bylaws, that first vote only concerns whether the assembly may debate it: members are voting on whether they'll be allowed to vote.

In Parliament, the bill known as "Darcos," after senator Laure Darcos, passed the French Senate unanimously in early April 2026. It would create a presumption that AI providers exploit cultural content, in other words reverse the burden of proof: the company would have to show it did not use a work, instead of the author having to prove it did. Sent on to the National Assembly, it was left off the agenda on 11 May, and the minister in charge of AI, Anne Le Hénanff, calls it "technically impossible to apply" and a source of "devastating legal risk."

In the background stands Anthropic. In September 2025, the American company behind the Claude assistant agreed to pay $1.5 billion to settle a copyright case. The roughly 500,000 works involved include books by Gaël Faye, Guillaume Musso, Amélie Nothomb, Joël Dicker and Albert Camus, and French-language authors had until 23 March 2026 to check whether they could claim compensation.

These three cases share the same anger, and I think they aim badly, because they lump together situations the law already tells apart.

What the American courts actually punished

The Anthropic case has become a slogan, "AI plundered authors," and the slogan has a basis. It still summarizes poorly what the judge decided, and the detail matters.

In June 2025, Judge William Alsup of the federal court for the Northern District of California issued a ruling in two parts, and each side quotes one while forgetting the other.

In the first, he held that training a model on lawfully acquired books is "fair use," the American doctrine that allows certain uses without permission. He even called the technology "exceedingly transformative," and AI's defenders wave that phrase around as a blanket absolution.

In the second, quoted far less often, he held that downloading pirated copies from shadow libraries such as LibGen or PiLiMi (illegal repositories of scanned books) to build a permanent library falls outside fair use. On that point, Alsup refused to side with Anthropic and sent the case to trial on damages. The $1.5 billion settlement therefore concerns how Anthropic obtained the books, and leaves the question of learning aside.

Read the ruling and the line is clear: the court punished Anthropic for stealing the books, while the learning itself fell under fair use. It's the difference between buying a book and slipping it under your coat, except that this shelf held millions of titles.

Three ways a book gets into a model

Today's debate annoys me on both sides. AI's opponents talk about training as a single, always guilty category; its defenders talk about "transformative use" as if it settled everything, including how the books were obtained. Both avoid the question that separates the cases: did the work enter the model with its author's consent?

The first case is training without consent or payment, on books scraped from LibGen. That is a scandal, and authors are right to go to court.

The second is training with agreement and payment. In November 2024, HarperCollins, one of the five big American publishers, signed a licensing deal with Microsoft covering its nonfiction backlist: $5,000 per title, split equally with the author, for three years, with individual opt-in required and safeguards (the model cannot reproduce more than 200 consecutive words or 5% of a book). An author who signs this with full knowledge is selling a license. You may find the price indecent, and I'll come back to that, but nobody stole from them.

The third is training excluded by default. Online writing apps such as WriteControl, Extypis and Scribbook have chosen not to feed any model with their users' texts. That path needs nobody's permission and is perfectly valid.

Mixing these three cases amounts to confusing theft, sale and abstention. By treating every AI provider as a presumed exploiter, the Darcos bill puts the company that paid HarperCollins on the same footing as the one that scraped LibGen, and it punishes the seller and the thief alike. I understand its intent, which is to spare authors a proof they could never supply. A presumption that doesn't separate lawful acquisition from plunder still erases the line that makes copyright defensible, and it's probably no accident that the minister in charge finds it unworkable herself.

Europe has already chosen consent

The American trial tends to drown this out, but "fair use" doesn't exist in European law. In Europe, the question is framed in terms of an explicit refusal, what lawyers call the opt-out.

The 2019 European copyright directive (known as the "DSM" directive, for the digital single market) created an exception for text and data mining, or "TDM," the automated analysis of large volumes of works. This exception allows training as long as the rights holder has not objected in an explicit, machine-readable way. European law has therefore already made consent the pivot of the system, by leaving authors the possibility of saying no.

The Darcos bill sits on top of this regime, and its real value lies in making that refusal enforceable, giving authors a weapon when a company claims it used nothing. Seen this way, the useful fight is about making the refusal effective and the proof attainable, which decides in practice whom you sue, and for what.

The price is a separate fight, and a legitimate one

That leaves the question that divides even the supporters of consent: $3,000 per book, is that serious? American writers have already sued in California, targeting training practices and also the settlement mechanisms themselves, which they consider tailored to the companies. They have good reasons to, since three thousand dollars for a book that took three years to write, set against models valued in the tens of billions, looks more like a wholesale rate than compensation.

An unfair price remains a problem separate from the principle. A badly paid license is a bad negotiation, and it doesn't make licensing illegitimate. Confuse the two and you hand AI's defenders a gift, since they only need to answer on the amount to look as if they had answered everything else. Authors have every interest in demanding a better license, and the fight over the amount is fair as long as it doesn't become an excuse to reject any market at all.

The same confusion on the writing side

The same muddle shows up on the neighboring question, "will AI write our books for us?"

Developers adopted AI within two years, and nobody claims they stopped being developers, since they reject most of what the tool suggests and decide on every line. The problem starts when 280 pages come out of a prompt and someone signs the cover, what I call, after others, vibe writing, modeled on the "vibe coding" that serious developers already scorn. A writer who asks for a synonym is getting help, while one who publishes a generated book is being replaced. The line runs in the same place as for training: what counts is the author's consent and involvement, whatever the tool.

In both debates, realities that stand opposed get filed under a single word, "training" or "AI that writes," and as long as they're confused, the fight goes badly.

The question to ask

I don't blame the SACD for wanting to sue, because it's right about the pirates, and a little fighting spirit does no harm. I only note that a collecting society accusing models of "aggregating without transparency" is itself, like all its peers, regularly asked by its own members to make clearer what it pays out to smaller authors. Transparency is a good fight, and it applies to those leading it too.

For each model, the useful question is whether it stole the books or paid for them. As long as the law, authors' societies and editorials keep asking whether to attack AI in general, they'll offer plunderers the shelter of confusion, because a thief has every interest in talk about training in general, where their case drowns in that of the neighbor who paid.

So I'm for consented, paid training, against stolen training, and I defend the right of anyone who wishes to refuse everything on their own turf. Keeping those three positions apart protects authors far better than merging them into one indignation.

---

Sources: Judge William Alsup's ruling in Bartz v. Anthropic, 23 June 2025 (Goodwin, Authors Alliance, Norton Rose Fulbright); Anthropic's $1.5 billion settlement, September 2025, and author compensation, deadline 23 March 2026 (ActuaLitté, ADAGP); the Darcos bill, passed by the French Senate in early April 2026 and left off the National Assembly agenda on 11 May 2026 (Sénat, ADAGP, SNAC, Décideurs Magazine); resolution by SACD members for the 25 June 2026 general assembly (ActuaLitté); HarperCollins and Microsoft deal, November 2024, $5,000 per title (Authors Guild, Publishers Weekly, Jane Friedman); Directive (EU) 2019/790, text and data mining exception and opt-out (arts. 3 and 4).